CUSTOMER CLUSTERING TO PERFORM A MARKETING CAMPAIGN¶
This project notebook is about customer clustering for an ad marketing campaign. I used different machine learning techniques such as k-means model, Principal Component Analysis(PCA) and autoencoders to perform dimensionality reduction.
UNDERSTANDING THE PROBLEM STATEMENT AND BUSINESS CASE¶
DATASET AND FEATURES¶
- Note: MSRP is the manufacturer's suggested retail price(MSRP) or sticker price. It respresents the suggested retal price of products. MRSP is used to standardize the price of products over multiple store location
WALKTHROUGH¶
In [89]:
import pandas as pd
import numpy as np
import seaborn as sns
import matplotlib.pyplot as plt
import zipfile
#import cv2
import plotly.express as px
import tensorflow as tf
from tensorflow.python.keras import Sequential
from tensorflow.keras import layers, optimizers
from tensorflow.keras.applications.resnet50 import ResNet50
from tensorflow.keras.layers import Input, Add, Dense, Activation, ZeroPadding2D, BatchNormalization, Flatten, Conv2D, AveragePooling2D, MaxPooling2D, Dropout
from tensorflow.keras.models import Model, load_model
from tensorflow.keras.initializers import glorot_uniform
from tensorflow.keras.utils import plot_model
from tensorflow.keras.callbacks import ReduceLROnPlateau, EarlyStopping, ModelCheckpoint, LearningRateScheduler
from IPython.display import display
from tensorflow.keras import backend as K
from sklearn.preprocessing import StandardScaler, normalize
from sklearn.cluster import KMeans
from sklearn.decomposition import PCA
from sklearn.metrics import silhouette_score
from sklearn.manifold import TSNE
from sklearn.metrics.pairwise import cosine_similarity
import plotly.express as px
import plotly.offline as py
import plotly.graph_objects as go
#from google.colab import files #library to upload files to colab notebook
%matplotlib inline
In [90]:
#loading my dataset
sales_df=pd.read_csv('sales_data_sample.csv', encoding='unicode_escape')
In [101]:
sales_df
Out[101]:
In [15]:
sales_df.dtypes
Out[15]:
In [8]:
sales_df['ORDERDATE']
Out[8]:
In [91]:
#convert ORDERDATE into datetime pandas format
sales_df['ORDERDATE']=pd.to_datetime(sales_df['ORDERDATE'])
In [92]:
#Null elements in 'ADDRESSLINE2'
pd.isna(sales_df['ADDRESSLINE2']).sum()
Out[92]:
Information concerning the types and number of non-null elements for each variable
In [8]:
sales_df.info()
Number of non-null elements for each variable in the dataset
In [7]:
sales_df.isna().sum() #we can also use 'isnull().sum()'
Out[7]:
Since there are lots of null values in ADDRESSLINE2, STATE, POSTALCODE and TERRITORY, we decide to drop them. We have several options to take into account graphical information. We can choose to keep COUNTRY or CITY or both, we choose to keep country. Plus, variables such us names of costumers are dropped since not relevant here.
In [93]:
df_drop=['ADDRESSLINE1','ADDRESSLINE2','POSTALCODE','TERRITORY','PHONE','STATE','CONTACTFIRSTNAME','CONTACTLASTNAME','CUSTOMERNAME','ORDERNUMBER']
sales_df=sales_df.drop(df_drop, axis=1)
sales_df.head()
Out[93]:
In [6]:
sales_df.isnull().sum() # Non-null values in the dataset
Out[6]:
In [7]:
# How many unique values of country, product code and product line
sales_df.nunique()
Out[7]:
EXPLORATORY DATA ANALYSIS AND DATA CLEANING¶
- Demography of customers
In [34]:
sales_df['COUNTRY'].value_counts().index # Customer's countries
Out[34]:
In [35]:
sales_df['COUNTRY'].value_counts() # Number of customer in each country
Out[35]:
- Small function: This function take into argument one feature such as COUNTRY and plot a barplot to the visualise the number of items.
In [103]:
def barplot_visualization(x):
fig=plt.Figure(figsize=(12,6))
fig=px.bar(x=sales_df[x].value_counts().index,y=sales_df[x].value_counts(),color=sales_df[x].value_counts().index,height=600)
fig.show(renderer="notebook")
we call te barplot_visualisation function and we pass it COUNTRY as an argument to visualise number of items per country.
In [105]:
barplot_visualization('COUNTRY')
In [106]:
barplot_visualization('CITY') # Visualisation of number of customer per city
- Our dataset is unbalanced considering CITY and COUNTRY. The number of Customers from USA really overpass the rest. Let's have a look on STATUS.
In [39]:
sales_df['STATUS'].nunique() # number of unique values for STATUS
Out[39]:
In [107]:
barplot_visualization('STATUS')
- Almost all transactions are shipped, the dataset extremely unbalanced regarding the STATUS
- A solution to this could be to downsampling the shipped class or upsampling the classes other than shipped class. Here we're simply going to drop the STATUS feature.
In [41]:
#drop the status variable
sales_df.drop(columns=['STATUS'], inplace=True)
sales_df.head()
Out[41]:
In [42]:
barplot_visualization('PRODUCTLINE')
- We define again a small function to add dummy variables to replace categorical variables in our dataset
In [43]:
def dummies(x):
dummy=pd.get_dummies(sales_df[x])
sales_df.drop(columns=x, inplace=True)
return pd.concat([sales_df,dummy], axis=1)
- Applying the above function to add dummy variables for countries in our dataset
In [44]:
sales_df=dummies('COUNTRY')
sales_df.head()
Out[44]:
Obtaining dummies for PRODUCTLINE and DEALSIZE
In [45]:
#Obtain dummies for 'PRODUCTLINE' and 'DEALSIZE'
sales_df=dummies('PRODUCTLINE')
sales_df=dummies('DEALSIZE')
Since the number of unique product code is 109, if we add one-hot variables, there would be additional 109 columns, we can avoid that by doing categorical encoding. It's important to avoid curse of dimensionality.
In [46]:
sales_df['PRODUCTCODE']=pd.Categorical(sales_df['PRODUCTCODE']).codes
- We're then going to aggregate sales to determine when(month) generaly the sales have a peak.
In [48]:
# grouping data by order date
sales_df_group=sales_df.groupby(by="ORDERDATE").sum()
sales_df_group
Out[48]:
In [49]:
#using a line plot
fig=plt.Figure(figsize=(12,6))
fig=px.line(x=sales_df_group.index,y=sales_df_group['SALES'],height=600)
fig.show() # we observe peak of the sales mainly in november
- Evaluating feature correlations
In [50]:
# correlation matrix between variables
plt.figure(figsize=(20,20))
corr_matrix=sales_df.iloc[:,:10].corr()
sns.heatmap(corr_matrix, annot=True, cbar=False)
Out[50]:
It looks like the Quarter ID and the monthly IDs are highly correlated Then lets's drop 'QTR_ID'. Then we can observe a strong correlation between MSRP and PRICEEACH. It is normal because MSRP are standardized prices. and PRICEEACH and SALES are also highly correlated, we can understand that for when PRICEEACH goes up SALES also goes up. SALES are obviously PRICEEACH multiplied by QUANTITY.
In [51]:
sales_df.drop("QTR_ID", inplace=True, axis=1)
sales_df.shape
Out[51]:
Let's plot distplots: Distplot shows the (1) histogram, (2) kde plot and (3) rug plot.
- (2) Kde Plot: Kernel Density Estimate is used for visualizing the Probability Density of a continuous variable.
- (3) Rug plot: plot of data for a single quantitative variable, displayed as marks along an axis (one-dimensional scatter plot).
In [52]:
import plotly.figure_factory as ff
plt.figure(figsize=(10,10))
for i in range(8):
if sales_df.columns[i] !='ORDERLINENUMBER' and sales_df.columns[i] !='ORDERDATE':
fig=ff.create_distplot([sales_df[sales_df.columns[i]].apply(lambda x: float(x))], ['distplot'])
fig.update_layout(title_text=sales_df.columns[i])
fig.show()
- QUANTITYORDERED are sort of uniformly distributed
In [53]:
sales_df.drop('ORDERDATE', axis=1, inplace=True)
Let's Visualize the relationship between variables using pairplots.
In [54]:
%matplotlib inline
plt.figure(figsize=(15,15))
fig=px.scatter_matrix(sales_df, dimensions=sales_df.columns[:8], color='MONTH_ID')
fig.update_layout(
title='Sales Data',
width=1100,
height=1100,
)
fig.show()
- A trend exists between 'SALES' and 'QUANTITYORDERED'
- A trend exists between 'MSRP' and 'PRICEEACH'
- A trend exists between 'PRICEEACH' and 'SALES'
- It seems that sales growth exists as we move from 2013 to 2014 to 2015 ('SALES' vs. 'YEAR_ID')
After performing this data exploratory we move to clustering of our data. We are going to use the elbow method to determine the optimal number of clusters(a minimum of three group) first.
- The elbow method is a heuristic method of interpretation and validation of consistency within cluster analysis designed to help find the appropriate number of clusters in a dataset.
- The idea is that Weighted Som of Squares(WSS) of the distances
- If the line chart looks like an arm, then the "elbow" on the arm is the value of k (number of clusters) that is the best.
UNDERSTANDING THE INTUITION AND THE THEORY BEHIND KMEANS CLUSTERING¶
FINDING THE OPTIMAL NUMBER OF CLUSTERS USING ELBOW METHOD¶
- The idea is the Within Cluster Sum of Squares (WCSS) goes down with the incresaing of the number of clusters.
- Up to k number of clusters the WCSS gets low enough and reach the "elbow" point, so the corresponding k is optimal
In [55]:
import copy
sales_df_copy=copy.copy(sales_df)
sales_df_copy.drop("CITY", axis=1, inplace=True) # we drop the variable 'CITY'
- Before searching the optimal k let's scale our data, it is necessary for the performance of our model
In [56]:
scaler=StandardScaler()
sales_df_copy_scaled=scaler.fit_transform(sales_df_copy)
In [57]:
sales_df_copy_scaled.shape
Out[57]:
Function to implement the elbow method
In [58]:
def elbow(k,X):
scores=[]
range_values=range(1,k)
for i in range_values:
kmeans=KMeans(n_clusters=i)
kmeans.fit(X)
scores.append(kmeans.inertia_) # inertia is the sum of the disances of samples to their cluster center
plt.plot(scores,'bx-')
plt.title('Finding right number of clusters')
plt.xlabel('clusters')
plt.ylabel('scores')
plt.show()
In [60]:
#calling the above function to determine the optimal number of clusters k
elbow(12,sales_df_copy_scaled)
- Notice the difficulty to determine the optimal number of cluster because the "elbow" is a little tricky to be identified from the curve, may be k=3 or k=5, not really clear right!
- One way to solution this is to implement a autoencoder, a dimensionality reduction model.But let's take first k=5 to examine clusters.
In [61]:
# Cluster the data using kmeans, k=5
kmeans=KMeans(5)
kmeans.fit(sales_df_copy_scaled)
labels=kmeans.labels_ #labels of clusters
In [53]:
labels
Out[53]:
In [54]:
kmeans.cluster_centers_.shape
Out[54]:
In [62]:
#Converting clusters centers into a dataframe
cluster_centers=pd.DataFrame(data=kmeans.cluster_centers_, columns=[sales_df.columns.drop('CITY')])
cluster_centers
Out[62]:
In [63]:
# In order to understand what these numbers mean, let's perform inverse transformation
cluster_centers = scaler.inverse_transform(cluster_centers)
cluster_centers = pd.DataFrame(data = cluster_centers, columns = [sales_df.columns.drop('CITY')])
cluster_centers
Out[63]:
Cluster 4 (Highest) - This group represents customers who buy items in high quantity centered around ~47, they buy items in all price range leaning towards high price items of ~99. They also correspond to the highest total sales around ~8293 and they are active throughout the year. They are the highest buyers of products with high MSRP ~158.
Cluster 3 - This group represents customers who buy items in varying quantity ~37, they tend to buy high price items ~95. Their sales is bit better average ~4442, they buy products with second highest MSRP of ~117.
Cluster 0 (lowest) - This group represents customers who buy items in low quantity ~30. They tend to buy low price items ~68. Their sales ~2030 is lower than other clusters and they are extremely active around holiday season. They buy products with low MSRP ~77.
- Cluster 2 - This group represents customers who are only active during the holidays. they buy in lower quantity ~34, but they tend to buy average price items around ~83. They also correspond to lower total sales around ~3055, they tend to buy items with MSRP around 86.
- Cluster 1 - This group represents customers who buy items in varying quantity ~32, they tend to buy average price items ~81. Their sales ~2991.
In [64]:
#Adding the labels on my data(on corresponding samples)
sales_df_cluster=pd.concat([sales_df_copy, pd.DataFrame({'cluster':labels})], axis=1)
sales_df_cluster
Out[64]:
A function to plot the histogram for each feature based on clusters
In [65]:
def hist_by_cluster(name_var, data,k):
'''A function to plot the histogram for each feature based on clusters'''
for i in name_var:
plt.figure(figsize=(30,6))
for j in range(k):
plt.subplot(1,k,j+1)
cluster=data[data['cluster']==j]
cluster[i].hist()
plt.title('{} \nCluster - {}'.format(i,j))
plt.show()
In [66]:
#Calling the above function
hist_by_cluster(sales_df_cluster.columns[:8],sales_df_cluster,5)
- Cluster-1 and cluster-2 are almost similar regarding sales, price of items and quantity ordered
- cluster-3 and cluster-4 are almost similar regarding the year and month of transactions and the quantity ordered
Reduction and visualization of the original data to 3 dimensional using PCA¶
In [67]:
#reduction of original data to 3 dimensional using PCA
pca=PCA(n_components=3)
principal_comp=pca.fit_transform(sales_df_copy_scaled)
principal_comp
Out[67]:
In [68]:
#Converting the PCA axes into a dataframe
pca_df=pd.DataFrame(data=principal_comp, columns=['pca1','pca2','pca3'])
pca_df.head()
Out[68]:
In [69]:
#concatenate the clusters labels to the dataframe
pca_df=pd.concat([pca_df,pd.DataFrame({'cluster':labels})], axis=1)
pca_df
Out[69]:
In [70]:
#Visualizing the clusters using 3D-Scatterplot
plt.figure(figsize=(30,30))
fig=px.scatter_3d(pca_df, x='pca1',y='pca2', z='pca3',color='cluster', symbol='cluster', size_max=18, opacity=0.7)
fig.update_layout(margin=dict(l=0,r=0,b=0,t=0))
Visualizing the clusters using 2D-Scatterplot: axe1 and axe2
In [71]:
plt.figure(figsize=(30,30))
fig=px.scatter(pca_df, x='pca1',y='pca2',color='cluster', symbol='cluster', size_max=18, opacity=0.7)
fig.update_layout(margin=dict(l=0,r=0,b=0,t=0))
- In the first graph the groups are not quite clear.
- In 2D It appears 2 group of customers we can consider, the rest are shared between the two. We have to solution this issues. After let's apply autoencoder to have clearer clusters.
PERFORM DIMENSIONALITY REDUCTION USING AUTOENCODERS¶
In [46]:
sales_df_copy.shape
Out[46]:
In [72]:
# Defining our autoencoder
from keras.optimizers import SGD
# Glorot Uniform initializer: https://keras.rstudio.com/reference/initializer_glorot_uniform.html
input_df=Input(shape=(37,))
x=Dense(50, activation='relu')(input_df)
x=Dense(500, activation='relu', kernel_initializer='glorot_uniform')(x)
x=Dense(500, activation='relu', kernel_initializer='glorot_uniform')(x)
x=Dense(2000, activation='relu', kernel_initializer='glorot_uniform')(x)
encoded=Dense(8, activation='relu',kernel_initializer='glorot_uniform')(x) #my encoder
x=x=Dense(500, activation='relu', kernel_initializer='glorot_uniform')(encoded)
decoded=Dense(37, activation='relu', kernel_initializer='glorot_uniform')(x) #my decoder
#autoendocder
autoencoder=Model(input_df, decoded)
#encoder - used for dimensionality reduction
encoder=Model(input_df, encoded)
autoencoder.compile(optimizer='adam',loss='mean_squared_error')
In [75]:
#Training our autoencoder
autoencoder.fit(sales_df_copy_scaled, sales_df_copy_scaled, batch_size=128, epochs=500, verbose=0)
Out[75]:
In [76]:
autoencoder.save_weights('autoencoder_1.h5')
In [77]:
pred=encoder.predict(sales_df_copy_scaled)# our dimension-reduced data
In [79]:
#Finding optimal number of clusters using our elbow function applied to our dimension-reduced data
elbow(12,pred)
Now It is clear that the elbow method above gives an optimal number of clusters k=3
In [80]:
# Cluster the data using kmeans, k=3
kmeans=KMeans(3)
kmeans.fit(sales_df_copy_scaled)
labels=kmeans.labels_
In [84]:
labels
Out[84]:
In [85]:
kmeans.cluster_centers_.shape
Out[85]:
In [81]:
#Converting cluster centers into a dataframe
cluster_centers=pd.DataFrame(data=kmeans.cluster_centers_, columns=[sales_df.columns.drop('CITY')])
cluster_centers
Out[81]:
In [82]:
# In order to understand what these numbers mean, let's perform inverse transformation
cluster_centers = scaler.inverse_transform(cluster_centers)
cluster_centers = pd.DataFrame(data = cluster_centers, columns = [sales_df.columns.drop('CITY')])
cluster_centers
Out[82]:
- Cluster 2 (High) - This group represents customers who buy items in high quantity centered around ~47, they buy items in range leaning towards high price items of ~99. They also correspond to the highest total sales around ~8293 and they are active throughout the year. They are the highest buyers of products with high MSRP ~158.
- Cluster 0 (middle) - This group represents customers who buy items in high quantity centered around ~37, they buy items in range leaning towards items of ~95. They correspond to the second total sales around ~4398. They are the buyers of products with MSRP ~115.
- Cluster 1 (low) - They represent the lowest groups in terms of sales ~ 2061, a MSRP around 77 and quantity centered around 69
In [83]:
#Adding the labels on my data(on corresponding samples)
sales_df_cluster=pd.concat([sales_df_copy, pd.DataFrame({'cluster':labels})], axis=1)
sales_df_cluster
Out[83]:
In [84]:
#Plot the histogram for each based on cluster
hist_by_cluster(sales_df_cluster.columns[:8],sales_df_cluster,3)
- Notice that the three clusters are really different along variables
Then let's perofrm PCA to visualize data along with the three(03) clusters
In [85]:
pca=PCA(n_components=3)
principal_comp=pca.fit_transform(sales_df_copy_scaled)
principal_comp
Out[85]:
In [86]:
pca_df=pd.DataFrame(data=principal_comp, columns=['pca1','pca2','pca3'])
pca_df.head()
Out[86]:
In [87]:
#concatenate the clusters labels to the dataframe
pca_df=pd.concat([pca_df,pd.DataFrame({'cluster':labels})], axis=1)
pca_df
Out[87]:
In [88]:
#Visualizing the clusters using 3D-Scatterplot
plt.figure(figsize=(30,30))
fig=px.scatter_3d(pca_df, x='pca1',y='pca2', z='pca3',color='cluster', symbol='cluster', size_max=18, opacity=0.7)
fig.update_layout(margin=dict(l=0,r=0,b=0,t=0))
- Excellent!! We have now three main clear clusters